Synthetic data enables human-grade microtubule analysis with foundation models for segmentation

Area of Science:

• Cell Biology
• Biophysics
• Microscopy Imaging

Background:

• Microtubules (MTs) are crucial for intracellular transport, cell division, and drug mechanisms.
• Manual segmentation of MTs is time-consuming and hinders systematic evaluation of automated methods.
• A lack of large-scale labeled datasets limits the assessment of current MT analysis tools.

Purpose of the Study:

• To address the need for labeled data by creating a synthetic dataset for MT segmentation.
• To benchmark existing automated methods and identify high-performing models for MT analysis.
• To demonstrate the feasibility of automated MT segmentation using synthetic data-guided AI.

Main Methods:

• Generation of the SynthMT synthetic dataset using a novel pipeline on interference reflection microscopy (IRM) data.
• Evaluation of nine automated MT analysis methods in zero- and few-shot settings.
• Hyperparameter Optimization (HPO) of the SAM3 model using the SynthMT dataset.

Main Results:

• Classical algorithms and current foundation models showed limitations in segmenting in vitro MTs from IRM images.
• The SAM3Text model, after HPO on SynthMT data, achieved near-perfect, sometimes superhuman, performance on real MT images.
• Synthetic data generation and HPO significantly improved automated MT segmentation accuracy.

Conclusions:

• Automated segmentation of microtubules is now feasible with AI models effectively guided by synthetic data.
• The SynthMT dataset, generation pipeline, and evaluation framework are publicly released to foster research progress.
• This work overcomes the bottleneck of manual annotation for MT analysis, enabling systematic method evaluation.